Mainul Hossain

Third year PhD student with 7 peer reviewed publications in software engineering, Human Computer Interaction, Artificial Intelligence, and Machine Learning. My research focuses on LLM utilization and usability in the software documentation paradigm, and I am seeking an internship to apply my skills and contribute to innovative projects with real world impact.

Ph.D. Research: ActiveLLM (Summer 2024 Ph.D. Research Award), 2023 - Present, 

• Researched and evaluated LLMs (ChatGPT, Claude, Llama) to detect, validate, and modify design rules in codebases through natural language interactions.
• Developed ActiveLLM, integrating LLMs with ActiveDocumentation for dynamic rule generation and real-time feedback within IDEs.
• Conducted a developer survey showing increased productivity by accelerating coding tasks and improving alignment with design rationale.

Quantitative currency evaluation in low-resource settings through pattern analysis to assist visually impaired users (RDM 2025 Workshop, IEEE ICDM 2025)

• Built an integrated three stream pipeline for Bangladeshi banknotes that performs denomination classification, usability assessment, and counterfeit detection in a single workflow for visually impaired users.
• Curated a unified 10 class dataset of 82,512 images and implemented perceptual hash based deduplication (Hamming threshold 5), then trained a lightweight Custom CNN (287K params, about 1.1 MB) achieving 0.9342 F1 on the combined dataset and 0.99 accuracy on counterfeit detection.
• Proposed the Unified Currency Damage Index (UCDI), a unified usability score (0 to 1) that fuses binary loss, RGB degradation, edge and corner integrity, symbolic feature loss, and damage clustering, with a rule based cap under severe symbolic loss.

STRIP: Supervised Training with Reinforcement Learning for Irrelevant Prompt Removal in Mathematical Reasoning

• Proposed STRIP, a two phase framework that trains an LLM to remove irrelevant sentences from noisy math word problems, generating a cleaned prompt in a single forward pass.
• Implemented STRIP on LLaMA 2 7B using LoRA adapters (r=64, alpha=128) with 8 bit quantization, and trained on GSM IC using supervised fine tuning followed by Direct Preference Optimization.
• Designed a multi component reward function with chunk level filtering rewards and asymmetric penalties (false negatives penalized more than false positives), plus faithfulness and completeness signals, achieving 0.9927 filtering F1 with SFT and 0.9894 with DPO on GSM IC.

Bulk Invoice Download and Export (Stripe Marketplace App), 2025, 

• Developed and published a Stripe Marketplace application enabling bulk downloading and exporting of multiple invoice PDFs directly within the Stripe Dashboard, significantly reducing manual effort for audits and tax preparations.
• Built with ReactJS (Frontend), Django, PostgreSQL, Celery, Redis, and containerized via Docker Compose; deployed on AWS Provides no-code integration visible across all Stripe Dashboard pages and listed on the Stripe Apps Marketplace.

Analyzing Negative Experiences in Hajj Pilgrimage, IEEE Access Vol 9 & Heliyon Vol 9 Issue 5, 

• Conducted large-scale surveys (online: n=236, in-person: n=752, total: n=988).
• Quantified integrity of data collectors using conflict and learning-based analyses, identifying up to 99% unreliable collectors and highlighting key religious and socio-geographical factors.
• Analyzed negative experiences using clustering (7 clusters, 21 issue types) and developed a tripartite graph to link issues with recommendations.
• Findings underline the necessity of human-in-the-loop validation and provide actionable insights for Hajj management personnel.